Surya for OCR and document layout

How it works

Surya exposes document-understanding tasks that can be used in a larger conversion pipeline. The geometry and recognized text are useful when you need to inspect what the OCR actually saw.

What you need

Pick the recognition and layout components needed for your pages. Capture geometry when debugging the processing of columns, table rows or lines with unusual placement.

What to try first

Run one low-resolution scan and one clean page. Compare line boxes and recognition output, then measure performance on documents in the expected scripts.

Save the bounding boxes with the text for several poor scans. That makes it possible to trace a recognition problem to missing lines or an incorrect text span.

Where to look for failures

OCR confidence and a clean-looking page can conceal individual character errors. Review model licensing and verify identifiers and numeric values with source images.

Related Franklin pages

Run the same scan through each approach and inspect the failed regions. An OCR engine and a full document-conversion pipeline can differ substantially in reading order and structural output.

• Detect and recognize text and extract structure with PaddleOCR pipelines. PaddleOCR for configurable document recognition
• Recognize printed text with a local engine and language data files. Tesseract for conventional local OCR

For how recognizing characters fits into the larger document processing workflow, see Franklin’s glossary. Optical character recognition

Project documentation

• Current project README
• README at the reviewed revision